DEV.co
Open-Source LLM · cyankiwi

Nex-N2-mini-AWQ-INT4

Nex-N2-mini is a 37B parameter quantized (INT4 AWQ) agentic language model built on Qwen3.5-35B, designed for tool calling, code generation, and long-horizon task execution. It is Apache 2.0 licensed, ungated, and available for immediate deployment. The model demonstrates strong performance on agent benchmarks and coding tasks, with calibration on STEM and agentic datasets. This is a quantized derivative of the base Nex-N2-mini model, optimized for reduced memory footprint and inference speed while retaining multi-lingual support (10 languages).

Source: HuggingFace — huggingface.co/cyankiwi/Nex-N2-mini-AWQ-INT4
37.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
108.7k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developercyankiwi
Parameters37.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads108.7k
Likes10
Last updated2026-07-03
Sourcecyankiwi/Nex-N2-mini-AWQ-INT4

What Nex-N2-mini-AWQ-INT4 is

Nex-N2-mini-AWQ-INT4 is a 4-bit quantized version of the Nex-N2-mini model using Activation-Aware Quantization (AWQ). Version 26.05.01. Base model: nex-agi/Nex-N2-mini (built on Qwen3.5-35B-A3B-Base). Model size on disk: 24.43 GB. Calibration data: STEM and Agentic datasets. Supported via HuggingFace endpoints and compressed-tensors format. Last modified: 2026-07-03. 108,667 downloads, 10 likes. Integrates with customized sglang fork for optimal serving. Context length and specific inference latency/throughput are not specified in the card.

Quickstart

Run Nex-N2-mini-AWQ-INT4 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="cyankiwi/Nex-N2-mini-AWQ-INT4")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Autonomous code agents and software engineering workflows

Strong performance on SWE-Bench Pro (50.2), Terminal-Bench 2.1 (60.7), and SWE-Bench Verified (74.4). Suitable for automated code review, bug fixing, and CI/CD integration where the model can iterate on feedback from test execution.

Agentic workflows with tool integration (search, APIs, terminal commands)

Benchmarked on tool calling (Toolathlon 33.3, WideSearch 62.0), real-world agent orchestration, and multi-step planning. Good fit for autonomous research, web scraping workflows, and systems that require reliable tool execution chains.

On-premise deployment in resource-constrained environments

24.43 GB model size and INT4 quantization makes this suitable for single or dual-GPU deployments (2–4× H100 configurations noted in card). Useful where latency and cost matter more than maximum quality, or for edge/private-infrastructure scenarios.

Running & fine-tuning it

Estimated 25–30 GB GPU VRAM for INT4-quantized model (exact VRAM depends on batch size and context length). Card references deployment on: (1) two 8×H100 servers (16 GPUs, tp=16) for Nex-N2-Pro; (2) one 2×H100 server (2 GPUs, tp=2) for Nex-N2-mini. A single H100 (80GB VRAM) or dual A100 (160GB total) should handle inference with these parameters. Docker image (nexagi/sglang:v0.5.12) requires at least 32GB shared memory (--shm-size 32g noted). No CPU-only inference details provided.

Card does not explicitly document LoRA, QLoRA, or other fine-tuning methodology. The model is post-trained on Qwen3.5 base; further instruction-tuning or domain adaptation capability is unknown. Recommend testing on small dataset or reaching out to nex-agi ([email protected]) for guidance on fine-tuning workflows, especially given INT4 quantization may complicate gradient-based tuning.

When to avoid it — and what to weigh

  • Highest-precision reasoning or safety-critical applications — INT4 quantization may reduce reasoning fidelity compared to full-precision. For medical, legal, or safety-critical inference where small errors compound, consider full-precision or compare degradation empirically.
  • Very long context windows (>100k tokens) with tight latency SLA — Context length is unknown from the card. If your workload requires massive context retention with sub-second latency, confirm specs with the publisher or test empirically.
  • Specialized domain models or fine-tuning on proprietary corpora — This is a general-purpose agentic model. Domain-specific fine-tuning feasibility is unknown; if you need guaranteed domain adaptation or IP-locked weights, evaluate LoRA/QLoRA capability first.
  • Setups requiring official commercial support or SLAs — Published as open source by nex-agi with contact email ([email protected]). No official SLA, support tier, or commercial indemnification stated. For production use, confirm support terms directly with nex-agi.

License & commercial use

Apache License 2.0 (apache-2.0). This is an OSI-approved, permissive open-source license allowing commercial use, modification, and redistribution with attribution and without warranty.

Apache 2.0 permits commercial use without restriction. Model is ungated (gated=false). However, commercial support, maintenance, and liability are not addressed in the card. For production deployments, review nex-agi's Terms of Service (if any) at nex-agi.com or contact [email protected] to confirm liability, indemnification, and support expectations. No commercial use restrictions detected, but no official commercial SLA is documented.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No explicit security audit, adversarial robustness testing, or safety alignment details provided in the card. As a quantized agentic model with tool-calling and code-execution capabilities, ensure: (1) input validation and sandboxing for tool calls; (2) monitoring for prompt injection or jailbreak attempts; (3) review of training data (calibration set: STEM and Agentic); (4) testing for model-generated code safety before execution in production. Quantization may change model behavior; validate on security-sensitive benchmarks before deployment.

Alternatives to consider

DeepSeek-V4-Pro

Comparable agentic and coding performance (e.g., 72.0 on Terminal-Bench 2.1 vs. Nex-N2-mini's 60.7). Open source, but may have different license restrictions; requires review.

Qwen3.5-32B (base model, unquantized)

Same Qwen base as Nex-N2-mini but without nex-agi's agentic post-training. Lower memory overhead with quantization but no adaptive thinking framework. Use if agentic features not required.

Nex-N2-Pro (full-precision or higher precision variant)

Stronger benchmarks (e.g., 75.3 on Terminal-Bench 2.1, 80.8 on SWE-Bench Verified) but larger footprint (~397B base). Use if latency/cost constraints allow and max quality is prioritized.

Software development agency

Ship Nex-N2-mini-AWQ-INT4 with senior software developers

Nex-N2-mini-AWQ-INT4 is Apache 2.0 licensed and ready for immediate download. Fork the GitHub repo, use our sglang Docker image, or deploy via HuggingFace Endpoints. Test on your hardware first, then contact nex-agi ([email protected]) for production support details.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Nex-N2-mini-AWQ-INT4 FAQ

Can I use this model commercially without restriction?
Yes, Apache 2.0 license permits commercial use. However, the license includes no warranty or liability protection. Verify with nex-agi ([email protected]) whether they provide commercial support, indemnification, or service-level agreements before deploying to production.
What GPU hardware is needed to run Nex-N2-mini locally?
Estimated 25–30 GB VRAM. Card shows examples on 2×H100 (tp=2). A single H100 or dual A100 should work. Exact throughput/latency depends on batch size and context length (latter not specified). Test on target hardware before production.
Is this model fine-tunable, and does it support LoRA?
Card does not document fine-tuning methodology or LoRA support. Contact nex-agi directly or test on a small dataset. INT4 quantization may complicate gradient-based fine-tuning.
What is the context length, and how does it compare to other models?
Unknown. Card does not specify maximum context window. Check the base Nex-N2-mini GitHub repo or contact nex-agi for this detail, as it affects suitability for long-horizon tasks.

Software development & web development with DEV.co

Adopting Nex-N2-mini-AWQ-INT4 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy an Agentic Code Assistant?

Nex-N2-mini-AWQ-INT4 is Apache 2.0 licensed and ready for immediate download. Fork the GitHub repo, use our sglang Docker image, or deploy via HuggingFace Endpoints. Test on your hardware first, then contact nex-agi ([email protected]) for production support details.